Free Implementing Cisco Data Center AI Infrastructure Practice Questions

Sample Practice Questions

Try these 10 sample questions from the Implementing Cisco Data Center AI Infrastructure practice bank. Each includes the answer and explanation. Continue with the interactive practice above to review the full question set.

1. 1What does retrieval-augmented generation add to a model's response-generation process?

   1. A. External information retrieved to inform the response
   2. B. Parameter updates made during every user request
   3. C. New model weights learned from an unlabeled corpus
   4. D. A compressed copy of the model's existing parameters

   Answer: A

   Explanation: Cisco states, "Using retrieval augmented generation (RAG), models can generate responses using external and current data." RAG supplies external information for generation; it is not itself a weight-training or compression method.

   Reference: FlashStack for Generative AI Inferencing Design Guide — Generative AI Opportunities > Extract Knowledge From Proprietary Data

2. 2An AI POD training deployment must keep GPU synchronization traffic isolated from enterprise ingress traffic. Which network design matches Cisco's dual-fabric architecture?

   1. A. Carry synchronization and user ingress through the same frontend network
   2. B. Place GPU synchronization on the backend and enterprise ingress on the frontend
   3. C. Carry training-data ingestion on an externally accessible backend network
   4. D. Place user ingress on the backend and reserve the frontend for GPU collectives

   Answer: B

   Explanation: Cisco states, "AI training and fine-tuning workloads, and the associated GPU clusters, typically rely on two independent networks: a backend (BE) for parallel processing and synchronization across GPU nodes, and a frontend (FE) network for ingesting training data, checkpointing, and logging (Figure 8)." It also notes backend fabrics are isolated. Separating these roles meets the requested design; reversing or combining them does not.

   Reference: Cisco AI POD for Enterprise Training and Fine-Tuning Design Guide — Design > Dual Fabric Design

3. 3A Nexus fabric must pause a congested RoCE traffic class while leaving unrelated best-effort classes able to transmit. Which flow-control configuration supports that behavior?

   1. A. Enable link-level pause for every traffic class together
   2. B. Configure PFC for the designated no-drop traffic class
   3. C. Make the RoCE class drop eligible when its buffer fills
   4. D. Use ECN marking alone without enabling class-specific pause

   Answer: B

   Explanation: Cisco states, "PFC can signal congestion and prevents packet loss by pausing traffic for specific classes." PFC supplies the requested class pause. Link-wide pause stops unrelated classes, dropping loses traffic, and ECN marking alone signals congestion without providing that pause behavior.

   Reference: Cisco AI POD for Enterprise Training and Fine-Tuning Design Guide — Design > Quality of Service > Priority Flow Control

4. 4Two infrastructure configurations will run the same complete training job with identical model, dataset, and convergence criteria. The objective is to compare how quickly each finishes the job. Which measurement should be collected?

   1. A. The average GPU utilization percentage during the training run
   2. B. The average duration of an individual training iteration
   3. C. The achieved bandwidth of individual inter-GPU collective operations
   4. D. The elapsed wall-clock time from training-job start to completion

   Answer: D

   Explanation: Cisco states, "The primary performance metric for these workloads is Job Completion Time (JCT)." For the identical complete training jobs specified, start-to-completion time directly measures how quickly each finishes. Utilization, iteration duration, and collective bandwidth diagnose components but are not the complete job duration.

   Reference: Cisco AI POD for Enterprise Training and Fine-Tuning Design Guide — AI Workloads and Infrastructure Implications > Training and Fine-Tuning Workloads

5. 5Which description correctly characterizes how model training uses prepared datasets?

   1. A. Applying fixed learned behavior to answer a live request
   2. B. Learning patterns and relationships from prepared datasets
   3. C. Retrieving documents to supplement an existing model's answer
   4. D. Measuring response latency without changing model behavior

   Answer: B

   Explanation: Cisco states, "Prepared datasets feed into high-performance training environments where models learn patterns and relationships." Training learns from data; serving, retrieval, and measurement describe different activities.

   Reference: What is an AI factory? — How an AI factory works > 2. Model training and tuning

6. 6One distributed training job produces low-entropy flows, and several equal-cost links remain underused despite sufficient aggregate capacity. What should the network evaluation prioritize?

   1. A. Increasing the number of equal-cost paths without reviewing flow placement
   2. B. Selecting switches only by their aggregate port bandwidth rating
   3. C. Moving the storage dataset without examining inter-GPU traffic
   4. D. Assessing load-balancing behavior for the workload's low flow entropy

   Answer: D

   Explanation: Cisco states, "AI training workloads typically have low entropy, which can make traditional load balancing methods like ECMP sub-optimal." Aggregate capacity alone therefore does not establish effective utilization; evaluation must examine how this traffic is distributed.

   Reference: Cisco AI POD for Enterprise Training and Fine-Tuning Design Guide — Design > Backend Fabric Design > Key Requirements > Efficient Load Balancing

7. 7A Nexus RoCE egress queue is configured for ECN marking as occupancy rises. What should happen when congestion marking is triggered for ECN-capable packets?

   1. A. Leave the ECN field unchanged and rely only on packet-loss detection
   2. B. Discard each selected packet instead of delivering a congestion mark
   3. C. Replace end-to-end marking with a link-wide Ethernet pause operation
   4. D. Set the IP ECN field to signal experienced congestion to the receiver

   Answer: D

   Explanation: Cisco states, "packets are randomly marked to indicate Congestion Experienced using the ECN bits in the IP/TOS header." The configured behavior marks ECN-capable packets in the IP header. Unchanged fields, loss-based notification, or a link-wide pause are different congestion responses.

   Reference: Cisco AI POD for Enterprise Training and Fine-Tuning Design Guide — Design > Quality of Service > Queuing with Congestion Avoidance

8. 8An inference benchmark records a prompt at 10.0 seconds, the first output token at 10.7 seconds, and the completed response at 13.0 seconds. Which value is time-to-first-token?

   1. A. 0.7 seconds
   2. B. 2.3 seconds
   3. C. 3.0 seconds
   4. D. 10.7 seconds

   Answer: A

   Explanation: Cisco states, "Time-to-First Token (TTFT): Measures the interval between a user prompt to receiving and the first output token." Subtract the prompt timestamp from the first-token timestamp: 10.7 − 10.0 = 0.7 seconds. Completion delay and absolute timestamps are different measurements.

   Reference: Cisco AI POD for Enterprise Training and Fine-Tuning Design Guide — AI Workloads and Infrastructure Implications > Inferencing Workloads > Time-to-First Token

9. 9A trained generative model receives a prompt and produces a new response using its learned knowledge. Which workload is being executed?

   1. A. Training from scratch
   2. B. Parameter fine-tuning
   3. C. Generative AI inference
   4. D. Training-data labeling

   Answer: C

   Explanation: Cisco defines the workload: "Generative AI inferencing refers to the process of using a trained generative AI model (large language models and non-large language models) to generate new data or content based on input or contextual cues." The trained model is being applied, rather than trained or supplied with labeled training data.

   Reference: FlashStack for Generative AI Inferencing Design Guide — What is Generative AI Inferencing?

10. 10Multiple AI tenants share a physical fabric but require separate virtual networks and routing contexts. Which architecture in the Cisco AI POD design addresses both requirements?

    1. A. VXLAN EVPN tenant networks with separate VRF instances
    2. B. One shared VLAN and one shared routing table for all tenants
    3. C. One common virtual network with traffic distinguished only by DSCP
    4. D. A common VRF with all tenant routes redistributed without isolation

    Answer: A

    Explanation: Cisco states, "VXLAN provides data plane isolation by encapsulating each tenant’s traffic into a unique virtual network (L2 or L3), while MP-BGP ensures control plane isolation through Virtual Routing and Forwarding (VRF) instances." Shared VLANs, shared VRFs, or DSCP marking alone do not supply both forms of tenant separation.

    Reference: Cisco AI POD for Enterprise Training and Fine-Tuning Design Guide — How AI PODs Address Enterprise AI Challenges > Multi-tenancy